Hierarchical Bidirected Graph Convolutions for Large-Scale 3-D Point Cloud Place Recognition

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3-D point cloud data offers robustness for place recognition compared to 2-D images.
• Extracting informative features from 3-D point clouds presents challenges due to difficulties in defining convolution operations.

Purpose of the Study:

• To propose a novel hierarchical bidirected graph convolution network (HiBi-GCN) for large-scale 3-D point cloud place recognition.
• To address the challenge of feature extraction in 3-D point cloud data.

Main Methods:

• Introduced a new hierarchical kernel defined as a hierarchical graph structure via unsupervised clustering.
• Employed pooling edges to aggregate hierarchical graphs from fine to coarse.
• Utilized fusing edges to combine pooled graphs from coarse to fine.

Main Results:

• The HiBi-GCN method learns representative features hierarchically and probabilistically.
• The approach extracts discriminative and informative global descriptors for place recognition.
• Experimental results confirm the suitability of the hierarchical graph structure for 3-D scene representation.

Conclusions:

• The proposed hierarchical graph structure is well-suited for point cloud-based place recognition.
• HiBi-GCN enhances the ability to represent and recognize real-world 3-D scenes effectively.